MSCA-PLA: multi-scale cross-attention with differentiable pooling for protein-ligand binding affinity prediction

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Protein-ligand binding affinity prediction is vital for rational drug design.
• Current methods struggle to capture synergistic effects within atomic clusters and fine-grained pairwise interactions simultaneously.
• Adaptive identification of key atomic clusters driving interactions remains a significant challenge.

Purpose of the Study:

• To develop a novel computational framework for accurate protein-ligand binding affinity prediction.
• To effectively model both multi-scale atomic interactions (clusters and pairwise contacts).
• To address the challenge of adaptively identifying binding-relevant atomic clusters.

Main Methods:

• Proposed MSCA-PLA: a multi-scale cross-attention framework with differentiable pooling.
• Adaptive aggregation of atom-level embeddings into binding-relevant clusters via differentiable pooling.
• Cluster-level and atom-level bidirectional cross-attention modules to model cooperative and pairwise interactions.
• Staged gated fusion strategy to integrate multi-scale interaction information.

Main Results:

• MSCA-PLA demonstrates effectiveness in predicting protein-ligand binding affinity.
• The framework successfully models cooperative interactions among atom groups and fine-grained pairwise contacts.
• Validation across benchmark datasets, diverse-protein settings, and virtual screening tasks confirms performance.
• Effectiveness supported on an external structural subset.

Conclusions:

• MSCA-PLA provides an effective approach for protein-ligand binding affinity prediction by integrating multi-scale interaction information.
• The framework's ability to model both cluster-level synergy and atom-level contacts enhances predictive power.
• This method holds promise for advancing drug discovery and virtual screening applications.